sglangv0.5.2 & support Qwen3-Next-80B-A3B-Instruct
This commit is contained in:
maxiao1
599
benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
Normal file
599
benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
Normal file
@@ -0,0 +1,599 @@
|
||||
# Adapted from https://github.com/vllm-project/vllm/blob/main/benchmarks/kernels/benchmark_moe.py
|
||||
import argparse
|
||||
import json
|
||||
import time
|
||||
from contextlib import nullcontext
|
||||
from datetime import datetime
|
||||
from typing import Any, Dict, List, Tuple, TypedDict
|
||||
|
||||
import ray
|
||||
import torch
|
||||
import triton
|
||||
from ray.experimental.tqdm_ray import tqdm
|
||||
from transformers import AutoConfig
|
||||
|
||||
from sglang.srt.layers.moe.fused_moe_triton import override_config
|
||||
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
|
||||
fused_moe,
|
||||
get_config_dtype_str,
|
||||
get_config_file_name,
|
||||
get_default_config,
|
||||
get_moe_configs,
|
||||
)
|
||||
from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
|
||||
from sglang.srt.layers.moe.topk import TopKConfig, select_experts
|
||||
from sglang.srt.utils import is_hip
|
||||
|
||||
_is_hip = is_hip()
|
||||
|
||||
|
||||
class BenchmarkConfig(TypedDict):
|
||||
BLOCK_SIZE_M: int
|
||||
BLOCK_SIZE_N: int
|
||||
BLOCK_SIZE_K: int
|
||||
GROUP_SIZE_M: int
|
||||
num_warps: int
|
||||
num_stages: int
|
||||
|
||||
|
||||
def benchmark_config(
|
||||
config: BenchmarkConfig,
|
||||
num_tokens: int,
|
||||
num_experts: int,
|
||||
shard_intermediate_size: int,
|
||||
hidden_size: int,
|
||||
topk: int,
|
||||
dtype: torch.dtype,
|
||||
use_fp8_w8a8: bool,
|
||||
use_int8_w8a8: bool,
|
||||
use_int8_w8a16: bool,
|
||||
block_shape: List[int] = None,
|
||||
num_iters: int = 100,
|
||||
) -> float:
|
||||
init_dtype = torch.float16 if use_fp8_w8a8 else dtype
|
||||
x = torch.randn(num_tokens, hidden_size, dtype=dtype)
|
||||
if use_int8_w8a16 or use_int8_w8a8:
|
||||
w1 = torch.randint(
|
||||
-127,
|
||||
127,
|
||||
(
|
||||
num_experts,
|
||||
shard_intermediate_size,
|
||||
hidden_size,
|
||||
),
|
||||
dtype=torch.int8,
|
||||
)
|
||||
w2 = torch.randint(
|
||||
-127,
|
||||
127,
|
||||
(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
shard_intermediate_size // 2,
|
||||
),
|
||||
dtype=torch.int8,
|
||||
)
|
||||
else:
|
||||
w1 = torch.randn(
|
||||
num_experts, shard_intermediate_size, hidden_size, dtype=init_dtype
|
||||
)
|
||||
w2 = torch.randn(
|
||||
num_experts, hidden_size, shard_intermediate_size // 2, dtype=init_dtype
|
||||
)
|
||||
gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32)
|
||||
|
||||
w1_scale = None
|
||||
w2_scale = None
|
||||
a1_scale = None
|
||||
a2_scale = None
|
||||
if use_int8_w8a16:
|
||||
w1_scale = torch.randn(
|
||||
(num_experts, 2 * shard_intermediate_size), dtype=torch.float32
|
||||
)
|
||||
w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
|
||||
if use_fp8_w8a8 or use_int8_w8a8:
|
||||
if use_int8_w8a8 and block_shape is None:
|
||||
w1_scale = torch.randn(
|
||||
num_experts, shard_intermediate_size, dtype=torch.float32
|
||||
)
|
||||
w2_scale = torch.randn(num_experts, hidden_size, dtype=torch.float32)
|
||||
elif block_shape is None:
|
||||
w1_scale = torch.randn(num_experts, dtype=torch.float32)
|
||||
w2_scale = torch.randn(num_experts, dtype=torch.float32)
|
||||
a1_scale = torch.randn(1, dtype=torch.float32)
|
||||
a2_scale = torch.randn(1, dtype=torch.float32)
|
||||
else:
|
||||
block_n, block_k = block_shape[0], block_shape[1]
|
||||
n_tiles_w1 = (shard_intermediate_size + block_n - 1) // block_n
|
||||
n_tiles_w2 = (hidden_size + block_n - 1) // block_n
|
||||
k_tiles_w1 = (hidden_size + block_k - 1) // block_k
|
||||
k_tiles_w2 = (shard_intermediate_size // 2 + block_k - 1) // block_k
|
||||
w1_scale = torch.rand(
|
||||
(num_experts, n_tiles_w1, k_tiles_w1), dtype=torch.float32
|
||||
)
|
||||
w2_scale = torch.rand(
|
||||
(num_experts, n_tiles_w2, k_tiles_w2), dtype=torch.float32
|
||||
)
|
||||
|
||||
if use_fp8_w8a8:
|
||||
w1 = w1.to(torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn)
|
||||
w2 = w2.to(torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn)
|
||||
|
||||
input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32)
|
||||
topk_config = TopKConfig(
|
||||
top_k=topk,
|
||||
renormalize=True,
|
||||
)
|
||||
topk_output = select_experts(x, input_gating, topk_config)
|
||||
|
||||
def prepare(i: int):
|
||||
input_gating = gating_output[i]
|
||||
new_topk_output = select_experts(x, input_gating, topk_config)
|
||||
topk_output.topk_weights.copy_(new_topk_output.topk_weights)
|
||||
topk_output.topk_ids.copy_(new_topk_output.topk_ids)
|
||||
topk_output.router_logits.copy_(new_topk_output.router_logits)
|
||||
|
||||
def run():
|
||||
moe_runner_config = MoeRunnerConfig(
|
||||
inplace=True,
|
||||
)
|
||||
|
||||
with override_config(config):
|
||||
fused_moe(
|
||||
x,
|
||||
w1,
|
||||
w2,
|
||||
topk_output,
|
||||
moe_runner_config=moe_runner_config,
|
||||
use_fp8_w8a8=use_fp8_w8a8,
|
||||
use_int8_w8a8=use_int8_w8a8,
|
||||
use_int8_w8a16=use_int8_w8a16,
|
||||
w1_scale=w1_scale,
|
||||
w2_scale=w2_scale,
|
||||
a1_scale=a1_scale,
|
||||
a2_scale=a2_scale,
|
||||
block_shape=block_shape,
|
||||
)
|
||||
|
||||
# JIT compilation & warmup
|
||||
run()
|
||||
torch.cuda.synchronize()
|
||||
|
||||
# Capture 10 invocations with CUDA graph
|
||||
graph = torch.cuda.CUDAGraph()
|
||||
with torch.cuda.graph(graph):
|
||||
for _ in range(10):
|
||||
run()
|
||||
torch.cuda.synchronize()
|
||||
|
||||
# Warmup
|
||||
for _ in range(5):
|
||||
graph.replay()
|
||||
torch.cuda.synchronize()
|
||||
|
||||
start_event = torch.cuda.Event(enable_timing=True)
|
||||
end_event = torch.cuda.Event(enable_timing=True)
|
||||
|
||||
latencies: List[float] = []
|
||||
for i in range(num_iters):
|
||||
prepare(i)
|
||||
torch.cuda.synchronize()
|
||||
|
||||
start_event.record()
|
||||
graph.replay()
|
||||
end_event.record()
|
||||
end_event.synchronize()
|
||||
latencies.append(start_event.elapsed_time(end_event))
|
||||
avg = sum(latencies) / (num_iters * 10) * 1000 # us
|
||||
graph.reset()
|
||||
return avg
|
||||
|
||||
|
||||
def get_rocm_configs_compute_bound() -> List[Dict[str, int]]:
|
||||
configs: List[BenchmarkConfig] = []
|
||||
waves_per_eu_range = 0
|
||||
for num_stages in [2]:
|
||||
for block_m in [32, 64, 128, 256]:
|
||||
for block_k in [32, 64, 128, 256]:
|
||||
for block_n in [16, 32, 64, 128, 256]:
|
||||
for num_warps in [1, 2, 4, 8]:
|
||||
for group_size in [1, 4, 8, 16, 32]:
|
||||
configs.append(
|
||||
{
|
||||
"BLOCK_SIZE_M": block_m,
|
||||
"BLOCK_SIZE_N": block_n,
|
||||
"BLOCK_SIZE_K": block_k,
|
||||
"GROUP_SIZE_M": group_size,
|
||||
"num_warps": num_warps,
|
||||
"num_stages": num_stages,
|
||||
"waves_per_eu": waves_per_eu_range,
|
||||
}
|
||||
)
|
||||
return configs
|
||||
|
||||
|
||||
def get_configs_compute_bound() -> List[Dict[str, int]]:
|
||||
# Reduced search space for faster tuning.
|
||||
# TODO(woosuk): Increase the search space and use a performance model to
|
||||
# prune the search space.
|
||||
configs: List[BenchmarkConfig] = []
|
||||
if _is_hip:
|
||||
configs = get_rocm_configs_compute_bound()
|
||||
else:
|
||||
for num_stages in [2, 3, 4, 5]:
|
||||
for block_m in [16, 32, 64, 128, 256]:
|
||||
for block_k in [64, 128, 256]:
|
||||
for block_n in [32, 64, 128, 256]:
|
||||
for num_warps in [4, 8]:
|
||||
for group_size in [1, 16, 32, 64]:
|
||||
configs.append(
|
||||
{
|
||||
"BLOCK_SIZE_M": block_m,
|
||||
"BLOCK_SIZE_N": block_n,
|
||||
"BLOCK_SIZE_K": block_k,
|
||||
"GROUP_SIZE_M": group_size,
|
||||
"num_warps": num_warps,
|
||||
"num_stages": num_stages,
|
||||
}
|
||||
)
|
||||
return configs
|
||||
|
||||
|
||||
@ray.remote(num_gpus=1)
|
||||
class BenchmarkWorker:
|
||||
|
||||
def __init__(self, seed: int) -> None:
|
||||
torch.set_default_device("cuda")
|
||||
torch.cuda.manual_seed_all(0)
|
||||
self.seed = seed
|
||||
# Get the device ID to allocate tensors and kernels
|
||||
# on the respective GPU.
|
||||
self.device_id = int(ray.get_gpu_ids()[0])
|
||||
|
||||
def benchmark(
|
||||
self,
|
||||
num_tokens: int,
|
||||
num_experts: int,
|
||||
shard_intermediate_size: int,
|
||||
hidden_size: int,
|
||||
topk: int,
|
||||
dtype: torch.dtype,
|
||||
use_fp8_w8a8: bool,
|
||||
use_int8_w8a8: bool,
|
||||
use_int8_w8a16: bool,
|
||||
block_shape: List[int],
|
||||
) -> Tuple[Dict[str, int], float]:
|
||||
torch.cuda.manual_seed_all(0)
|
||||
dtype_str = get_config_dtype_str(
|
||||
dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
|
||||
)
|
||||
# NOTE(woosuk): The current naming convention uses w2.shape[2], which
|
||||
# is the intermediate size after silu_and_mul.
|
||||
block_n = block_shape[0] if block_shape else 0
|
||||
block_k = block_shape[1] if block_shape else 0
|
||||
op_config = get_moe_configs(
|
||||
num_experts, shard_intermediate_size // 2, dtype_str, block_n, block_k
|
||||
)
|
||||
if op_config is None:
|
||||
config = get_default_config(
|
||||
num_tokens,
|
||||
num_experts,
|
||||
shard_intermediate_size,
|
||||
hidden_size,
|
||||
topk,
|
||||
dtype_str,
|
||||
False,
|
||||
block_shape,
|
||||
)
|
||||
else:
|
||||
config = op_config[min(op_config.keys(), key=lambda x: abs(x - num_tokens))]
|
||||
with torch.cuda.device(self.device_id) if is_hip() else nullcontext():
|
||||
kernel_time = benchmark_config(
|
||||
config,
|
||||
num_tokens,
|
||||
num_experts,
|
||||
shard_intermediate_size,
|
||||
hidden_size,
|
||||
topk,
|
||||
dtype,
|
||||
use_fp8_w8a8,
|
||||
use_int8_w8a8,
|
||||
use_int8_w8a16,
|
||||
block_shape,
|
||||
)
|
||||
return config, kernel_time
|
||||
|
||||
def tune(
|
||||
self,
|
||||
num_tokens: int,
|
||||
num_experts: int,
|
||||
shard_intermediate_size: int,
|
||||
hidden_size: int,
|
||||
topk: int,
|
||||
dtype: torch.dtype,
|
||||
use_fp8_w8a8: bool,
|
||||
use_int8_w8a8: bool,
|
||||
use_int8_w8a16: bool,
|
||||
block_shape: List[int],
|
||||
search_space: List[Dict[str, int]],
|
||||
) -> Dict[str, int]:
|
||||
best_config = None
|
||||
best_time = float("inf")
|
||||
with torch.cuda.device(self.device_id) if is_hip() else nullcontext():
|
||||
for config in tqdm(search_space):
|
||||
try:
|
||||
kernel_time = benchmark_config(
|
||||
config,
|
||||
num_tokens,
|
||||
num_experts,
|
||||
shard_intermediate_size,
|
||||
hidden_size,
|
||||
topk,
|
||||
dtype,
|
||||
use_fp8_w8a8,
|
||||
use_int8_w8a8,
|
||||
use_int8_w8a16,
|
||||
block_shape,
|
||||
num_iters=10,
|
||||
)
|
||||
except (triton.runtime.autotuner.OutOfResources, RuntimeError):
|
||||
# Some configurations may be invalid and fail to compile.
|
||||
continue
|
||||
|
||||
if kernel_time < best_time:
|
||||
best_time = kernel_time
|
||||
best_config = config
|
||||
now = datetime.now()
|
||||
print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
|
||||
assert best_config is not None
|
||||
return best_config
|
||||
|
||||
|
||||
def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
|
||||
return {
|
||||
"BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
|
||||
"BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
|
||||
"BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
|
||||
"GROUP_SIZE_M": config["GROUP_SIZE_M"],
|
||||
"num_warps": config["num_warps"],
|
||||
"num_stages": config["num_stages"],
|
||||
**(
|
||||
{"waves_per_eu": config["waves_per_eu"]} if "waves_per_eu" in config else {}
|
||||
),
|
||||
}
|
||||
|
||||
|
||||
def save_configs(
|
||||
configs: Dict[int, BenchmarkConfig],
|
||||
num_experts: int,
|
||||
shard_intermediate_size: int,
|
||||
hidden_size: int,
|
||||
topk: int,
|
||||
dtype: torch.dtype,
|
||||
use_fp8_w8a8: bool,
|
||||
use_int8_w8a8: bool,
|
||||
use_int8_w8a16: bool,
|
||||
block_shape: List[int],
|
||||
) -> None:
|
||||
dtype_str = get_config_dtype_str(
|
||||
dtype,
|
||||
use_int8_w8a16=use_int8_w8a16,
|
||||
use_fp8_w8a8=use_fp8_w8a8,
|
||||
use_int8_w8a8=use_int8_w8a8,
|
||||
)
|
||||
|
||||
# NOTE(woosuk): The current naming convention uses w2.shape[2], which
|
||||
# is the intermediate size after silu_and_mul.
|
||||
filename = get_config_file_name(
|
||||
num_experts,
|
||||
shard_intermediate_size // 2,
|
||||
dtype_str,
|
||||
block_shape,
|
||||
)
|
||||
|
||||
print(f"Writing best config to {filename}...")
|
||||
with open(filename, "w") as f:
|
||||
json.dump(configs, f, indent=4)
|
||||
f.write("\n")
|
||||
|
||||
|
||||
def main(args: argparse.Namespace):
|
||||
print(args)
|
||||
|
||||
config = AutoConfig.from_pretrained(args.model, trust_remote_code=True)
|
||||
if config.architectures[0] == "DbrxForCausalLM":
|
||||
E = config.ffn_config.moe_num_experts
|
||||
topk = config.ffn_config.moe_top_k
|
||||
intermediate_size = config.ffn_config.ffn_hidden_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
elif config.architectures[0] == "JambaForCausalLM":
|
||||
E = config.num_experts
|
||||
topk = config.num_experts_per_tok
|
||||
intermediate_size = config.intermediate_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
elif config.architectures[0] in ["Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"]:
|
||||
E = config.num_experts
|
||||
topk = config.num_experts_per_tok
|
||||
intermediate_size = config.moe_intermediate_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
elif config.architectures[0] in ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]:
|
||||
E = (
|
||||
config.n_routed_experts + (0 if args.disable_shared_experts_fusion else 1)
|
||||
if config.architectures[0] in ["DeepseekV3ForCausalLM"]
|
||||
else config.n_routed_experts
|
||||
)
|
||||
topk = config.num_experts_per_tok
|
||||
intermediate_size = config.moe_intermediate_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
elif config.architectures[0] == "Llama4ForConditionalGeneration":
|
||||
E = config.text_config.num_local_experts + (
|
||||
0 if args.disable_shared_experts_fusion else 1
|
||||
)
|
||||
topk = config.text_config.num_experts_per_tok
|
||||
intermediate_size = config.text_config.intermediate_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
elif config.architectures[0] in [
|
||||
"Grok1ForCausalLM",
|
||||
"Grok1ImgGen",
|
||||
"Grok1AForCausalLM",
|
||||
]:
|
||||
E = config.num_local_experts
|
||||
topk = config.num_experts_per_tok
|
||||
intermediate_size = config.moe_intermediate_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
elif config.architectures[0] in ["Glm4MoeForCausalLM"]:
|
||||
E = config.n_routed_experts
|
||||
topk = config.num_experts_per_tok
|
||||
intermediate_size = config.moe_intermediate_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
else:
|
||||
# Default: Mixtral
|
||||
E = config.num_local_experts
|
||||
topk = config.num_experts_per_tok
|
||||
intermediate_size = config.intermediate_size
|
||||
shard_intermediate_size = 2 * intermediate_size // args.tp_size
|
||||
|
||||
hidden_size = getattr(config, "hidden_size", None) or config.text_config.hidden_size
|
||||
dtype = config.torch_dtype
|
||||
use_fp8_w8a8 = args.dtype == "fp8_w8a8"
|
||||
use_int8_w8a8 = args.dtype == "int8_w8a8"
|
||||
use_int8_w8a16 = args.dtype == "int8_w8a16"
|
||||
block_shape = None
|
||||
if (
|
||||
hasattr(config, "quantization_config")
|
||||
and "weight_block_size" in config.quantization_config
|
||||
):
|
||||
block_shape = config.quantization_config["weight_block_size"]
|
||||
assert len(block_shape) == 2
|
||||
|
||||
if args.batch_size is None:
|
||||
batch_sizes = [
|
||||
1,
|
||||
2,
|
||||
4,
|
||||
8,
|
||||
16,
|
||||
24,
|
||||
32,
|
||||
48,
|
||||
64,
|
||||
96,
|
||||
128,
|
||||
256,
|
||||
512,
|
||||
1024,
|
||||
1536,
|
||||
2048,
|
||||
3072,
|
||||
4096,
|
||||
]
|
||||
else:
|
||||
batch_sizes = [args.batch_size]
|
||||
|
||||
ray.init()
|
||||
num_gpus = int(ray.available_resources()["GPU"])
|
||||
workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
|
||||
|
||||
def _distribute(method: str, inputs: List[Any]) -> List[Any]:
|
||||
outputs = []
|
||||
worker_idx = 0
|
||||
for input_args in inputs:
|
||||
worker = workers[worker_idx]
|
||||
worker_method = getattr(worker, method)
|
||||
output = worker_method.remote(*input_args)
|
||||
outputs.append(output)
|
||||
worker_idx = (worker_idx + 1) % num_gpus
|
||||
return ray.get(outputs)
|
||||
|
||||
if args.tune:
|
||||
search_space = get_configs_compute_bound()
|
||||
if block_shape is not None:
|
||||
block_n, block_k = block_shape[0], block_shape[1]
|
||||
search_space = [
|
||||
config
|
||||
for config in search_space
|
||||
if block_k % config["BLOCK_SIZE_K"] == 0
|
||||
]
|
||||
print(f"Start tuning over {len(search_space)} configurations...")
|
||||
|
||||
start = time.perf_counter()
|
||||
configs = _distribute(
|
||||
"tune",
|
||||
[
|
||||
(
|
||||
batch_size,
|
||||
E,
|
||||
shard_intermediate_size,
|
||||
hidden_size,
|
||||
topk,
|
||||
dtype,
|
||||
use_fp8_w8a8,
|
||||
use_int8_w8a8,
|
||||
use_int8_w8a16,
|
||||
block_shape,
|
||||
search_space,
|
||||
)
|
||||
for batch_size in batch_sizes
|
||||
],
|
||||
)
|
||||
best_configs = {
|
||||
M: sort_config(config) for M, config in zip(batch_sizes, configs)
|
||||
}
|
||||
save_configs(
|
||||
best_configs,
|
||||
E,
|
||||
shard_intermediate_size,
|
||||
hidden_size,
|
||||
topk,
|
||||
dtype,
|
||||
use_fp8_w8a8,
|
||||
use_int8_w8a8,
|
||||
use_int8_w8a16,
|
||||
block_shape,
|
||||
)
|
||||
end = time.perf_counter()
|
||||
print(f"Tuning took {end - start:.2f} seconds")
|
||||
else:
|
||||
outputs = _distribute(
|
||||
"benchmark",
|
||||
[
|
||||
(
|
||||
batch_size,
|
||||
E,
|
||||
shard_intermediate_size,
|
||||
hidden_size,
|
||||
topk,
|
||||
dtype,
|
||||
use_fp8_w8a8,
|
||||
use_int8_w8a8,
|
||||
use_int8_w8a16,
|
||||
block_shape,
|
||||
)
|
||||
for batch_size in batch_sizes
|
||||
],
|
||||
)
|
||||
|
||||
for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
|
||||
print(f"Batch size: {batch_size}, config: {config}")
|
||||
print(f"Kernel time: {kernel_time:.2f} us")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
"--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1"
|
||||
)
|
||||
parser.add_argument("--tp-size", "--tp", type=int, default=2)
|
||||
parser.add_argument(
|
||||
"--dtype",
|
||||
type=str,
|
||||
choices=["auto", "fp8_w8a8", "int8_w8a16", "int8_w8a8"],
|
||||
default="auto",
|
||||
)
|
||||
parser.add_argument("--seed", type=int, default=0)
|
||||
parser.add_argument("--batch-size", type=int, required=False)
|
||||
parser.add_argument("--tune", action="store_true")
|
||||
parser.add_argument("--disable-shared-experts-fusion", action="store_true")
|
||||
args = parser.parse_args()
|
||||
|
||||
main(args)
|
||||
Reference in New Issue
Block a user